基于沙漏阶梯残差模型的胸部影像多标签分类-《中国医学物理学杂志》

文章信息/Info

Title:: Multi-label chest X-ray classification using sandglass ladder residual network

Author(s):: FANG Junze; XING Suxia; GUO Zheng; LI Kexian; WANG Yu; School of Artificial Intelligence, Beijng Technology and Business University, Beijng 100048, China

Keywords:: chest X-ray; multi-label classification; convolutional neural network; vision transformer; sandglass convolution

摘要:: 提出一种基于沙漏阶梯残差模型(SLRN),用于胸部影像疾病的多标签分类,提高临床诊断的准确性。SLRN 的设计包括 3 个关键模块,首先采用沙漏卷积模块同时提取通道间信息与空间信息;然后使用阶梯自注意力模块,通过移位操作实现不同窗口划分,扩大感受野,提取并融合多尺度特征;在多标签分类阶段,使用多头残差注意力,捕捉到不同标签之间的相关性和特征间的重要性,通过调整不同特征的权重实现更精准的分类。本研究在印第安纳大学收集的胸部 X 光数据集(IU X-Ray)和美国国立卫生研究院收集并公开的胸部 X 射线数据集(Chest X-Ray14)中进行验证,实验证明 SLRN 结合了卷积神经网络和视觉转换器的优点,可以捕捉影像中的局部特征和全局关联,更好地处理长距离依赖关系,辅助医生进行临床诊断。

Abstract:: A sandglass ladder residual network (SLRN) is proposed for multi-label chest X-ray classification, thereby improving the accuracy of clinical diagnosis. SLRN consists of 3 key modules: (1) a sandglass convolutional module to simultaneously extract inter-channel and spatial information; (2) a ladder self attention block to achieve different window divisions through shift operations, expand the receptive field, and realize multi-scale feature extraction and fusion; (3) class specific residual attention in the multi-label classification stage to capture the correlation between different labels and the importance of features for accomplishing more accurate classification by adjusting the weights of different features. The proposed model is validated using the IU X-Ray dataset collected by Indiana University and the publicly available Chest XRay14 dataset collected by the National Institutes of Health in the United States; and the results demonstrate that SLRN which combines the advantages of convolutional neural network and vision transformer enables the capture of local features and global correlations in images, better handles long-distance dependencies, and assists doctors in clinical diagnosis.